Sleep Headband Makers Harvest Millions of EEG Hours to Train AI

Wearable neurotechnology companies have quietly accumulated tens of millions of hours of electroencephalography recordings from paying customers and are using them to train AI models, a joint investigation by Science and the Pulitzer Center's AI Accountability Network has found. Several firms failed to disclose that use in their privacy policies before reporters started asking questions.

Interaxon, the Canadian company behind the Muse sleep and meditation headband, has already trained an AI model on roughly 50,000 hours of EEG recordings. The company says it strips identifying information, including names, email addresses and postal addresses, before the data enters the training pipeline. Chief Marketing Officer Nadia Kumentas was candid about the ambition: "We aim to use as much of that eligible corpus as possible."

What did customers actually agree to?

Until Science came calling, not much that covered this. Interaxon's privacy policy did not explicitly state that brain recordings could be used to train foundation models. The company has since updated the document, and Kumentas conceded the point: "We agree users should be explicitly informed."

The belated disclosure matters. Customers whose recordings were already consumed by model training received the classification after the fact, a sequence that legal exposure tends to follow. The consent problem is industry-wide: more than two-thirds of the companies examined explicitly allowed corporate transfers of data without user consent. Should a small neurotechnology firm be acquired, its archive of brain recordings would pass to a new owner by default. Colorado and California have amended their consumer privacy laws to cover neural data. Most other states have not.

Emotiv, another wearable EEG maker, is training its Large Brain model partly on recordings from customers using its high-end devices. It also updated its privacy policy after questions from Science, disclosing the use of de-identified neural data for model training subject to opt-in consent. Two policy revisions prompted by one investigation is a respectable yield.

How does a brain foundation model work?

The mechanics differ from a large language model. Instead of text, developers feed the model recordings of electrical activity in the brain. It learns recurring patterns that can be adapted to recognize different brain states or support research tasks.

The clinical results are real. AI startup Piramidal has trained its model on more than 4 million hours of EEG data in partnership with Cleveland Clinic. Its seizure-detection system correctly identified nine out of ten seizures in a pilot study after fine-tuning on the clinic's recordings.

The consumer hardware, however, is not a mind-reading machine. Most wearable EEG devices carry one to six electrodes, against the 19 to 25 found in clinical systems, and their recordings are correspondingly thinner. A consumer headband reads the weather; a clinical rig reads the forecast.

Is this the first neural data scare?

Not quite. In February, an engineer who bought a smart sleep mask discovered he could access other people's brainwave recordings. He began investigating after connection problems, examined the device's Android application, and found hardcoded credentials shared across devices. The credentials unlocked EEG readings belonging to other users.

The wider industry has been collecting neural data at scale for some time:

• Last December, San Francisco startup Coundit launched an effort to collect 10,000 hours of neural data from thousands of patients, using its own headset to capture brain activity alongside speech and text. The goal is AI that translates neural signals into language.
• Meta is developing BrainQwerty, a non-invasive brain-computer interface that converts brain activity into keyboard input. Its second-generation system raised average word accuracy from 40% to 61%. That is progress, and it remains far from practical daily use.

The commercial logic is straightforward. EEG archives are expensive to assemble and impossible to re-create retroactively, which makes them valuable precisely when consent frameworks are weakest. With two states regulating neural data and the rest silent, and with disclosure still arriving after the training run rather than before it, expect consent language and acquisition clauses in this market to come under renewed scrutiny, and possibly litigation, as the brain foundation model race accelerates.